lkml.org 
[lkml]   [2017]   [Nov]   [14]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: [PATCH RFC v3 1/6] x86/paravirt: Add pv_idle_ops to paravirt ops
From
Date


On 2017/11/14 18:27, Juergen Gross wrote:
> On 14/11/17 10:38, Quan Xu wrote:
>>
>> On 2017/11/14 15:30, Juergen Gross wrote:
>>> On 14/11/17 08:02, Quan Xu wrote:
>>>> On 2017/11/13 18:53, Juergen Gross wrote:
>>>>> On 13/11/17 11:06, Quan Xu wrote:
>>>>>> From: Quan Xu <quan.xu0@gmail.com>
>>>>>>
>>>>>> So far, pv_idle_ops.poll is the only ops for pv_idle. .poll is called
>>>>>> in idle path which will poll for a while before we enter the real idle
>>>>>> state.
>>>>>>
>>>>>> In virtualization, idle path includes several heavy operations
>>>>>> includes timer access(LAPIC timer or TSC deadline timer) which will
>>>>>> hurt performance especially for latency intensive workload like
>>>>>> message
>>>>>> passing task. The cost is mainly from the vmexit which is a hardware
>>>>>> context switch between virtual machine and hypervisor. Our solution is
>>>>>> to poll for a while and do not enter real idle path if we can get the
>>>>>> schedule event during polling.
>>>>>>
>>>>>> Poll may cause the CPU waste so we adopt a smart polling mechanism to
>>>>>> reduce the useless poll.
>>>>>>
>>>>>> Signed-off-by: Yang Zhang <yang.zhang.wz@gmail.com>
>>>>>> Signed-off-by: Quan Xu <quan.xu0@gmail.com>
>>>>>> Cc: Juergen Gross <jgross@suse.com>
>>>>>> Cc: Alok Kataria <akataria@vmware.com>
>>>>>> Cc: Rusty Russell <rusty@rustcorp.com.au>
>>>>>> Cc: Thomas Gleixner <tglx@linutronix.de>
>>>>>> Cc: Ingo Molnar <mingo@redhat.com>
>>>>>> Cc: "H. Peter Anvin" <hpa@zytor.com>
>>>>>> Cc: x86@kernel.org
>>>>>> Cc: virtualization@lists.linux-foundation.org
>>>>>> Cc: linux-kernel@vger.kernel.org
>>>>>> Cc: xen-devel@lists.xenproject.org
>>>>> Hmm, is the idle entry path really so critical to performance that a
>>>>> new
>>>>> pvops function is necessary?
>>>> Juergen, Here is the data we get when running benchmark netperf:
>>>>   1. w/o patch and disable kvm dynamic poll (halt_poll_ns=0):
>>>>      29031.6 bit/s -- 76.1 %CPU
>>>>
>>>>   2. w/ patch and disable kvm dynamic poll (halt_poll_ns=0):
>>>>      35787.7 bit/s -- 129.4 %CPU
>>>>
>>>>   3. w/ kvm dynamic poll:
>>>>      35735.6 bit/s -- 200.0 %CPU
>>>>
>>>>   4. w/patch and w/ kvm dynamic poll:
>>>>      42225.3 bit/s -- 198.7 %CPU
>>>>
>>>>   5. idle=poll
>>>>      37081.7 bit/s -- 998.1 %CPU
>>>>
>>>>
>>>>
>>>>   w/ this patch, we will improve performance by 23%.. even we could
>>>> improve
>>>>   performance by 45.4%, if we use w/patch and w/ kvm dynamic poll.
>>>> also the
>>>>   cost of CPU is much lower than 'idle=poll' case..
>>> I don't question the general idea. I just think pvops isn't the best way
>>> to implement it.
>>>
>>>>> Wouldn't a function pointer, maybe guarded
>>>>> by a static key, be enough? A further advantage would be that this
>>>>> would
>>>>> work on other architectures, too.
>>>> I assume this feature will be ported to other archs.. a new pvops makes
>>       sorry, a typo.. /other archs/other hypervisors/
>>       it refers hypervisor like Xen, HyperV and VMware)..
>>
>>>> code
>>>> clean and easy to maintain. also I tried to add it into existed pvops,
>>>> but it
>>>> doesn't match.
>>> You are aware that pvops is x86 only?
>> yes, I'm aware..
>>
>>> I really don't see the big difference in maintainability compared to the
>>> static key / function pointer variant:
>>>
>>> void (*guest_idle_poll_func)(void);
>>> struct static_key guest_idle_poll_key __read_mostly;
>>>
>>> static inline void guest_idle_poll(void)
>>> {
>>>     if (static_key_false(&guest_idle_poll_key))
>>>         guest_idle_poll_func();
>>> }
>>
>>
>> thank you for your sample code :)
>> I agree there is no big difference.. I think we are discussion for two
>> things:
>>  1) x86 VM on different hypervisors
>>  2) different archs VM on kvm hypervisor
>>
>> What I want to do is x86 VM on different hypervisors, such as kvm / xen
>> / hyperv ..
> Why limit the solution to x86 if the more general solution isn't
> harder?
>
> As you didn't give any reason why the pvops approach is better other
> than you don't care for non-x86 platforms you won't get an "Ack" from
> me for this patch.


It just looks a little odder to me. I understand you care about no-x86 arch.

Are you aware 'pv_time_ops' for arm64/arm/x86 archs, defined in
   - arch/arm64/include/asm/paravirt.h
   - arch/x86/include/asm/paravirt_types.h
   - arch/arm/include/asm/paravirt.h

I am unfamilar with arm code. IIUC, if you'd implement pv_idle_ops
for arm/arm64 arch, you'd define a same structure in
   - arch/arm64/include/asm/paravirt.h     or
   - arch/arm/include/asm/paravirt.h

.. instead of static key / fuction.

then implement a real function in
   - arch/arm/kernel/paravirt.c.

Also I wonder HOW/WHERE to define a static key/function, then to benifit
x86/no-x86 archs?

Quan
Alibaba Cloud

>>> And KVM would just need to set guest_idle_poll_func and enable the
>>> static key. Works on non-x86 architectures, too.
>>>
>> .. referred to 'pv_mmu_ops', HyperV and Xen can implement their own
>> functions for 'pv_mmu_ops'.
>> I think it is the same to pv_idle_ops.
>>
>> with above explaination, do you still think I need to define the static
>> key/function pointer variant?
>>
>> btw, any interest to port it to Xen HVM guest? :)
> Maybe. But this should work for Xen on ARM, too.
>
>
> Juergen
>

\
 
 \ /
  Last update: 2017-11-14 12:44    [W:0.058 / U:19.900 seconds]
©2003-2017 Jasper Spaans. hosted at Digital OceanAdvertise on this site